Wheat Yield Estimation Study Using Hyperspectral Vegetation Indices

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

Wheat yield estimation study using hyperspectral vegetation indices

·        The idea behind the Abstract introduction it’s not clear. The Authors should rewrite the first three lines to make it clearer. In addition, it was reported in the last paragraph the RMSE and MSE values without units. Therefore, these units should be included in the text after every value reported if applies.

·        Figure 1 can be improved if the coordinate or axis labels were bigger, especially in the mini maps inside the main plot. In this sense, the Authors could include more specific or technical information in the figure caption about the study area.

·        In Lines 89-90 it was stated that the vegetation indices were selected based on the normal wavelength ranges for wheat yield estimation. This wavelength ranges are important information that should be included explicitly in the text or a least in a Table to justify the proposed vegetation indices.

·        In Section “3.1. Modeling analysis and model evaluation” it is stated that that 2/3 of the experimental data were used for model development and the remaining for evaluation. The question is, why the Authors proposed this quantity? they experimented with other proportions?

·        The Authors should include in Table 2 the RMSE and MSE values for each vegetation index.

·        It is recommended that the information displayed in Table 2 should be sorted from minor to major based on the Pearson Correlation. The same recommendation applies to Table 3, but in this case based on the modeling accuracy.

Comments on the Quality of English Language

Minor editing of English language required.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

Comments and Suggestions for Authors

The article offers a significant and very interesting contribution to the field of agronomy, utilizing advanced remote sensing techniques to estimate wheat yield. However, limitations associated with sample size and the precision of certain models suggest that future research should focus on expanding experimental data and exploring alternative or adjusted methods to improve the accuracy of yield predictions. The interdisciplinary approach and the application of multiple methodologies are notable highlights, although refinement in the selection and application of models is necessary to achieve broader and more robust applicability. I reaffirm that the innovative idea made me read the article about four times. I based my analysis on the text and to contribute to the proposal.

Strong points of the manuscript:

- The use of hyperspectral data is an advanced approach that allows for a detailed analysis of the conditions of wheat plantations at different growth stages. This represents a significant improvement over traditional yield estimation methods, which are more time-consuming and error-prone.

- The study applies five different modeling methods - multiple linear regression, stepwise linear regression, principal component analysis, neural networks, and random forests - which allows for a comprehensive comparison of the effectiveness of these techniques in the specific context of agricultural yield prediction.

- The random forest method proved superior, with a validation R² of 0.946, indicating excellent predictive capability of the model. This technique also demonstrated robustness in terms of mean square error and cross-validation mean square error.

Weak Points and Inconsistencies of the Manuscript:

- The study only uses 30 experimental plots, which may not be representative of wheat yield under different geographical or management conditions. A larger sample size could provide more generalizable and reliable results.

- The models of multiple linear regression and principal component analysis showed lower modeling accuracies (R² of 0.199 and 0.422, respectively), indicating significant limitations of these methods for this type of data or application context.

- Although the methods of neural networks and random forests showed good accuracy, the validation methodology could be more robust to confirm the effectiveness of the models under variable conditions, which is not sufficiently detailed.

Methodological Errors or Inconsistencies:

- Although the study selects 27 vegetation indices based on their correlations with wheat yield, it is not entirely clear how this selection was made and whether the most relevant indices were chosen. A deeper analysis of the individual contribution of each index could improve understanding of their effectiveness.

- The comparison between different models is not uniform, as different software was used for implementation (SPSS for some and Python for others), which could influence the results due to differences in implementation algorithms. I suggest making this clear to the readers of the article.

- Were the data normally distributed for the Pearson correlation?

 

Things to Consider in the Results or to Explore in Another Manuscript:

- The article mentions the use of 27 different vegetation indices and presents the correlation of some of these indices with wheat yield. However, a deeper analysis of how each index individually contributes to the model's accuracy could be extremely valuable. This would include discussing which indices were most effective and why, as well as which indices showed weak or insignificant correlations and the possible reasons for this. This will contribute to issues of inconsistency pointed out in the text above and in the discussion.

- Although the article discusses the precision metrics of the models, such as R², RMSE, and MSE, an analysis of outliers and their influence on the final results could provide deeper insights. Understanding if certain observations or specific field conditions had a disproportionate impact on the models would help refine prediction techniques and identify areas that require specialized attention.

- Models with lower accuracy, such as multiple linear regression and principal component analysis, are mentioned, but a discussion of the implications of these low precision values for practical applications is necessary. For example, in what scenarios could these models still be useful? Are there specific conditions under which they might perform better?

- If the study captured data at different stages of wheat growth, an analysis of the temporal variation of these vegetation indices and their impact on yield estimation over time would be helpful. This would aid in better understanding how weather conditions and other environmental variables throughout the season affect the accuracy of the models.

- Although the article compares the performance of different modeling methods, a more detailed analysis of why certain methods (such as random forests and neural networks) outperform others in terms of accuracy would be useful. This could include a discussion on the non-linear nature of the data or the ability of different techniques to capture complexities in the relationship between vegetation indices and wheat yield.

Improvements in Discussions to Address Some Inconsistencies:

- The discussion could benefit from a more in-depth analysis of the specific limitations of each regression model used, especially those that exhibited lower accuracy. For instance, discussing why multiple linear regression and principal component analysis did not achieve as high accuracy as random forests could provide valuable insights into the suitability of these methods for hyperspectral data. The discussion could explore whether the nature of the vegetation indices, the choice of independent variables, or the model structure were contributing factors.

- Although the article briefly mentions other studies in the area of yield estimation using vegetation indices, a more detailed comparison with similar research could enrich the discussion. This would include an analysis of how the results of this study align or diverge from other works, and what innovations or methodological variations could explain these differences.

- The discussion could be expanded to consider the impact of climatic conditions and soil characteristics in the study areas, which are critical factors in agriculture and can significantly influence wheat yield. Including a discussion on how these factors interact with the selected vegetation indices and affect the accuracy of the prediction models could provide a more comprehensive understanding of the results.

- While the study provides a solid technical foundation, the discussion could be enriched with more practical recommendations for farmers and agricultural managers. Discussing how these models could be implemented in the daily management of crops, including data collection frequency and integration with other agricultural management practices, could increase the practical impact of the study.

- The discussion addresses future research directions but could detail more specific strategies to overcome the identified limitations. For example, exploring new vegetation indices, adjusting modeling algorithms, or utilizing emerging remote sensing technologies could be discussed in more depth.

 

Figures:

- Figure 1 presents a black square that we cannot clearly see what it represents. The quality needs to be improved.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 3 Report

Comments and Suggestions for Authors

1. 1. Hyperspectral camera hardware: 

Please gives more detailed of the portable hyperspectral camera used in data acquisition.

• Spectral Resolution, not sampling size 

• Spatial Resolution 

•  

• Image Acquisition Time 

• Image irradiance at the vegetations 

 

Hyperspectral imaging techniques vary, primarily categorized into snapshot and scanning-based modalities. For instance, two distinct types of imagers are discussed in the following studies: 

1. Snapshot hyperspectral imager 

https://doi.org/10.1364/OPTICA.440074 

2. Spectral scanning hyperspectral imager 

https://doi.org/10.1364/BOE.4.001486 

Please specify the type of hyperspectral imager utilized in this study for image acquisition. This clarification is crucial to understand the preprocessing required for images obtained through alternative imaging modalities. 

2. 2. What is the formular of reflectance R? Does the author subtract dark reflectance? This point needs clarification.  

3. 3. The manuscript mentions that data acquisition was conducted under clear-sky conditions with favorable lighting, but this description requires quantification. The impact of the light source on the hyperspectral spectrum is significant, and the manuscript should address the potential errors introduced by lighting conditions. It is suggested to explain the accuracy of the measurements under these conditions and discuss any possible compensation methods that could enhance measurement precision. 

4. 4. The computational complexity and noise robustness of the models used in the study should be elaborated. It is essential to specify the minimum sample size required for these models to yield accurate results. While the manuscript mentions using 30 data points, it should be assessed whether some models might require larger sample sizes for reliable modeling. This information will help in evaluating the robustness and applicability of the models in practical scenarios. 

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 2 Report

Comments and Suggestions for Authors

The article presents minimum qualifications to be published.